Philosophia Mathematica (III) (2005) 13, 44–60. doi:10.

1093/philmat/nki006

Learning from Questions on Categorical Foundations

Colin McLarty∗

We can learn from questions as well as from their answers. This paper
urges some things to learn from questions about categorical foundations
for mathematics raised by Geoffrey Hellman and from ones he invokes

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

from Solomon Feferman.

There are two ways to take the question ‘Does category theory provide a
framework for mathematical structuralism?’ (Hellman [2003]).1 In terms
of a working framework, Awodey had to say: ‘obviously, yes’ (Awodey
[2004], p. 54). Category theory has been the standard research framework
for topology, most algebra, and much functional analysis since the 1950s.2
It has been so in algebraic geometry and number theory since the 1960s
and increasingly in all mathematics. It has been the textbook method of
structuralist mathematics since the Algebras of Serge Lang [1965] and
then Mac Lane and Birkhoff [1967]. So Hellman went on to the theoretical
question whether ‘category theory provides an autonomous foundation for
mathematics as an alternative to set theory’ (Hellman [2003], p. 129).
I have said yes. Awodey [2004] is disinclined to any ‘foundation’—
though I think he gives the word an unduly strict meaning.3 But we do not

∗ Department of Philosophy, Case Western Reserve University, Cleveland, Ohio U. S. A.

44106 cxm7@po.cwru.edu.
1 This article is closely based on my part in a panel discussion organized by Steve
Awodey at the Chicago 2004 joint meeting of the ASL and the Midwest APA, with papers
by Geoffrey Hellman and myself, and comments by Dana Scott and Stewart Shapiro. I refer
often to Hellman’s part which was a work in progress and so not suitable for quotation. A
version of that talk is to appear as Hellman [to appear]. I cite ideas raised by Hellman’s
talk—to say they are important, and I thank Hellman for them—but not to guarantee they
represent his precise thinking then or now.
2 For historical accounts see, for example, discussion of the 1950s in Dieudonné [1981],
[1989], and all articles dealing with the time since 1945 in James [1999].
3 I take it that Russell showed no logical formulation can give what mathematics was
‘really all about in the first place’ as mathematics itself has changed from Babylonia to
today, see especially Russell [1919]. He and Zermelo found that no naïvely self-evident
axioms suffice for current mathematics. Independence proofs since then have shown they
were right. Aristotle asked for too much when he said axioms must not admit of any proof.
Any statement can be ‘proved’ from some other statement roughly as plausible. I follow the
line Saunders Mac Lane often takes ([1986], p. 406), whereby foundations are ‘proposals
for the organization of Mathematics’, which I believe is much like what Shapiro [1991]
means by ‘foundations without foundationalism’. To count as a foundation the axioms must

Philosophia Mathematica (III), Vol. 13 No. 1 © Oxford University Press, 2005, all rights reserved

Philma: “nki006” — 2005/1/21 — 10:31 — page 44 — #1

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 45

only learn from answers and their reasons. We also learn from questions,
and I want to urge some things to learn from Hellman’s questions, and from
the ones he invokes from Solomon Feferman’s ‘Categorical foundations
and foundations of category theory’ [1977].
Hellman emphasizes the question of how to tell the presupposition of
an axiom system, and the need to distinguish two senses of ‘structural
axioms’. In one sense ‘structural’ axioms posit entities with only structural
properties. In the other sense structural ‘axioms’ state structural properties
and posit no entities at all. Hellman well argues that any foundation for

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

mathematics must say what entities it posits and in what sense, so that
structural axioms in this second sense cannot be foundations. His argu-
ment builds on Feferman’s. The deepest point of Feferman’s paper as it
seems to me is to show that we want much more from a foundation than
formal adequacy and practical efficacy. In his metaphor, to accept a given
foundation merely because it is formally adequate and practically product-
ive is like ‘not needing to hear, once one has learned to compose music’
(Feferman [1977], p. 153). We want to hear the music.

1. Feferman
Feferman’s [1977] is the most sustained critique of categorical foundations
to date. Categorists consider his arguments well refuted. For example Bell
[1981] endorsed Feferman’s ideas provisionally and then, after pursuing
the subject for a time, he decided for categorical foundations in [1986],
[1988], and [1998]. Yet there has been no explicit reply to Feferman until
now, and it is worth giving because his position is more subtle than many
people realize. His critique occupies only five pages and can be summed
up in three points: Category theory cannot be a logical foundation; it is
also psychologically derivative; and it is unmusical. What are logically
and psychologically prior, he says, are notions of operation and collection.
He says categorical notions of arrows cannot be logically prior to set-
theoretic accounts of objects:
My use of ‘logical priority’ refers not to relative strength of
formal theories but to order of definition of concepts, in the

be independently plausible to a reasonably sophisticated mathematician. One requirement

of independence, for me, is that formal first-order versions of the axioms must suffice for
first-order proofs of the theorems. Categorical set theory is offered as a foundation for ‘all’
mathematics modulo the kind of constraints Gödel proved, as are some versions of axioms
for a category of categories. Elsewhere I discuss categorical foundations for specific parts of
mathematics, an idea I especially associate with William Lawvere, and which I have called
‘radically categorical foundations’ as in McLarty [1998]. So far as I know all categorical
foundationalists have enough faith in the advance of mathematics to say category theory
will not be the last word. It is the latest and currently best word in structuralist organization
of mathematics.

Philma: “nki006” — 2005/1/21 — 10:31 — page 45 — #2

46 McLARTY

cases where certain of these must be defined before others. For

example, the concept of vector space is logically prior to that
of linear transformation. (Feferman [1977], p. 152)
This brings us face to face with mathematical practice. The first math-
ematicians to work with linear transformations defined them as functions
on lists of numbers. They did not define ‘vector space’ at all, and at most
defined a ‘vector’ as either a ‘directed line segment’ or an ‘ordered triple
of numbers’. Those definitions of ‘vector’ are still used today by some

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

engineers and even a few physicists.
By the 1920s leading mathematicians, notably Hermann Weyl and John
von Neumann, knew this was not the best definition for their purposes.
For them and for many today the best order of definition was to define a
vector space as any commutative group acted on by a field of scalars. Then
a vector is any element of a vector space, which is to say anything can be a
vector by placing it in a suitable context, and nothing is in itself a vector.4
A linear transformation is a group homomorphism that preserves scalar
multiplication. Feferman uses this definition of linearity.
In turn, though, this is not the best definition for many purposes today.
The most widely used of Alexander Grothendieck’s ideas in practice is
his axiomatization of Abelian categories. This appeared in Grothendieck
[1957] twenty years before Feferman’s article. The most influential text-
book treatments today are Lang’s Algebra [1993], first printed in 1965,
and Hartshorne’s Algebraic Geometry [1977], published the same year
as Feferman’s article. On this definition a linear space is any object in
an Abelian category, although it is more normal to call the arrows of an
Abelian category linear and say little about the objects in the general theory.
An Abelian category with a suitable relation to a field k is a category of
k-vector spaces. On this view anything can be a vector space when placed
in a suitable categorical context, and nothing is in itself a vector space.
An Abelian category is a category of linear transformations between
linear objects. The axioms say nothing about those objects except that
they are domains and codomains of the transformations. In fact the arrows-
only formulation of the category axioms can axiomatize Abelian categories
without ever mentioning objects. Those axioms speak only of arrows, that
is, of transformations.
Linearity of the transformations is expressed by saying they add and
subtract, and composition preserves addition. So for any transformations

4 Admittedly, in Zermelo-Fraenkel set theory, vectors are not elements of vector spaces.
A vector space is a 6-tuple V , +V , k, +k , ·k , (.) of a set V with addition +V making
it a commutative group, and a set k with addition +k and multiplication ·k making it a
field, and a scalar multiplication (.). Vectors are elements of an element of a vector space
or something like that depending on exactly how you define 6-tuples. But the point is
clear.

Philma: “nki006” — 2005/1/21 — 10:31 — page 46 — #3

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 47

f, g, h :A→B and j, k :B→C:

f + g :A→B 0A,B :A→B −f :A→B

f +g =g+f (f + g) + h = f + (g + h)
f + 0A,B = f f + (−f ) = 0A,B
j (f + g) = (jf ) + (jg) (j + k)f = (jf ) + (kf )

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

An Abelian category satisfies these plus some axioms on products,
kernels, and quotients. See Grothendieck [1957], Lang [1993], Mac Lane
[1998], and for issues of axiomatics see especially Freyd [1964].
We can take these axioms alone as a categorical foundation for this
notion of linearity. Or we can interpret them in set theory to get a (much
stronger) set-theoretic foundation for the same notion. Either way the
axioms describe linear transformations without saying anything about vec-
tor spaces or other linear spaces—except, again, that those spaces are
domains and codomains of the transformations. And even that can be elim-
inated by using the arrows-only form of category theory. All the usual
theorems of linear algebra follow using transformations rather than ele-
ments. The resulting theory applies to structures much more general, and
more complicated from a set-theoretic point of view, than just the classical
vector spaces and modules.
In short, there is not just one notion of ‘vector’ or ‘linear transformation’,
and different ones have different formalizations. Others are described in the
appendices below. In formalizing any concept, if you want to achieve what
set-theoretic foundations achieve in the way they achieve it, then you need
set-theoretic foundations. Of course that could be categorical set theory.
If you want specifically what membership-based set-theoretic foundations
like ZF do, in the way they do it, although these specifics find no significant
echo in practice, then you need specifically membership-based set-theoretic
foundations like ZF. But those are not the only things you can do.
The point is that Feferman knows all this. He knows the Abelian cat-
egory axioms. He knows that these and related axioms and practices show
it is formally possible to make arrows prior, and it is productive in fact.5 It
remains a logical error in his sense despite people doing it successfully. It
is also a psychological error:

My claim above is that the general concepts of operation and

collection have logical priority with respect to structural notions
(such as ‘group’, ‘category’, etc.) because the latter are defined
5 Recall that at this stage in Feferman’s argument it is not a question of ultimate founda-
tions but of ‘order of definition of concepts. . . . [e.g.,] the concept of vector space is logically
prior to that of linear transformation’ (Feferman [1977], p. 152)

Philma: “nki006” — 2005/1/21 — 10:31 — page 47 — #4

48 McLARTY

in terms of the former but not conversely. At the same time, I

believe our experience demonstrates their psychological prior-
ity. I realize that workers in category theory are so at home in
their subject that they find it more natural to think in categor-
ical rather than set-theoretical terms, but I would liken this to
not needing to hear, once one has learned to compose music.
(Feferman [1977], p. 153)
Even in 1977 this method was hardly confined to category theorists.

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

The most famous exercise in Lang’s Algebra was on the Abelian cat-
egory axioms. After a ten-page introduction to homology built around these
axioms, the sole exercise read: ‘Take any book on homological algebra,
and prove all the theorems without looking at the proofs given in that book’
(Lang [1965], p. 105).6 Lang is not a category theorist and his Algebra has
been a standard graduate textbook for decades.
So a great many mathematicians miss the music Feferman cares for.
For him the music lies, almost, in the structure of the iterative hierarchy of
ZF sets, more in proof theory, and most of all in philosophic questions of
realism versus constructivism, which he wants to build into the foundations.
He poses an ambitious philosophical goal:
Since neither the realist nor constructivist point of view encom-
passes the other, there cannot be any present claim to a universal
foundation for mathematics, unless one takes the line of reject-
ing all that lies outside the favored scheme. Indeed, multiple
foundations in this sense may be necessary, in analogy to the
use of both wave and particle conceptions in physics. Moreover
it is conceivable that still other kinds of of theories of operations
and collections will be developed as a result of further experi-
ence and reflection. I believe that none of these considerations
affects the counter-thesis of this part, namely that founda-
tions for structural mathematics are to be sought in theories
of operations and collections (if they are to be sought at all).
(Feferman [1977], p. 151)
He does not know what the new theories may look like but he knows
what they should not look like:
To avoid misunderstanding, let me repeat that I am not
arguing for accepting current set-theoretical foundations of
mathematics. Rather, it is that on the platonist view of
6 This disappeared from later editions since by the mid 1970s essentially all homolo-
gical algebra books were themselves organized around those axioms and proofs, even if
the theorems were not explicitly stated in that generality. I.e., they had all taken Lang’s
advice.

Philma: “nki006” — 2005/1/21 — 10:31 — page 48 — #5

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 49

mathematics something like present systems of set theory must

be prior to any categorical foundations. (Feferman, p. 15)
Feferman offers a non-extensional theory of his own as progress towards
a correct non-platonist foundation for category theory. It starts with a frag-
ment of first-order Peano arithmetic; so it has a term 0, and an operator 
satisfying axioms:

x   = 0;

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

x  = y  → x = y.

Reading the obvious way, these say 0 is not a successor, and if the successor
of x is the successor of y then x = y. But there is no induction axiom at
this level, terms are not taken to be natural numbers, and  is not taken
to be successor. Feferman adds a partial application symbol so each term
becomes a partially defined operator. The formula xy  z says, intuitively,
‘applying operator x to argument y gives value z . The idea is familiar
from recursive-function theory when natural numbers are taken as codes
for partial recursive functions and then xy  z says the xth partial recursive
function is defined at argument y, and takes value z.
The theory handles collections by way of partial classifications defined
as partially defined operators whose values are all 0 or 1, where 1 is officially
written 0 . The idea is also familiar from recursive function theory, where
a number n is taken as coding a set S of numbers if n codes a classifying
operator for S in this sense:
(1) whenever nx is defined, then nx = 0 or nx = 1;
(2) nx = 1 if and only if x ∈ S.
Curiously, then, operators are prior to collections in the foundations of
Feferman’s theory. Apparently this order of definition will be reversed by
the time the system is applied to concepts like linear transformation and
vector space but Feferman’s article does not go that far.
Obviously I agree with Feferman that foundations of mathematics
should lie in a general theory of operations and collections, only I say the
currently best general theory of those calls them arrows and objects. It is cat-
egory theory. And I think there is no question of whether to seek foundations
for structural mathematics. Of course we should. The theoretical unity
and practical power of modern structural methods make them, to my ear,
actually finer music than proof theory or realism versus constructivism.
The efficacy of structuralism in practice makes it a more compelling topic
for foundations, to me. But Feferman asks the right question—the ques-
tion of whether we ‘hear the music’. It is not a matter of merely technical
logic.

Philma: “nki006” — 2005/1/21 — 10:31 — page 49 — #6

50 McLARTY

2. Hellman
Hellman’s paper takes up a number of themes from Feferman’s and some I
think are just wrong. It is a mistake when Geoffrey follows Feferman [1977]
to say: ‘There is frank acknowledgement that the notion of function is pre-
supposed, at least informally, in formulating category theory’ (Helmann
[2003], p. 133). A very general notion of function, older than set theory,
certainly does motivate category theory.7 But motivation is not presupposi-
tion. Feferman’s system is motivated by the idea of natural numbers coding

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

partial recursive functions. That does not mean his system presupposes
codings or recursive function theory.
My [2004] dealt with other concerns Hellman raised, notably whether
categorical set theory uses something so arcane as sheaves on topological
spaces in order to define the real numbers (it does not), and whether it can
express the replacement axiom scheme (it can, in terms closer to Cantor’s
than to Zermelo-Fraenkel) (McLarty [2004], pp. 38–41, 47–50).
Now Hellman has expanded on the really central issues though. They
concern recognizing the presuppositions of a theory and especially the
presuppositions of structural theories. Of course we cannot give the pre-
suppositions of ‘category theory’ per se because there are too many things
‘category theory’ can mean. It would be wrong to assume that in practice
‘category’ and ‘functor’ are usually algebraic or structural in the sense of
being axiomatic primitives with no intended interpretation.
When checked recently, forty-one of the latest fifty references to
‘category’ in Mathematical Reviews were to specific categories, i.e., they
did have intended interpretations—insofar as anything in mathematics
ever does. These uses of ‘category’ and ‘functor’ are as specific, as
un-‘algebraic’, as upper-level mathematics ever gets. They are meant as
already situated within a foundation and not themselves open to founda-
tional re-interpretation. Even if we go with Awodey’s idea of mathematics
as schematic, these papers take the ambient structure as a datum and not
an explicit variable of interest to their work.
In general category theory, the terms ‘category’ and ‘functor’ are taken
algebraically, in Hellman’s sense. They are general and do not have an
intended interpretation. This accounts for seven of the fifty latest Mathemat-
ical Reviews references. The two remaining references were to Philosophia
Mathematica papers we are discussing, Awodey’s [2004] and my [2004].

7 The motivation was more general than set-theoretic functions in that already in 1945
prominent examples included ‘measurable functions’—each of which is an equivalence
class of set-theoretic functions, where two functions are equivalent if they agree on a set of
measure 0. See Stone [1932], p. 23 and passim. Another example was ‘rational functions’ in
algebraic geometry—each one of which is an equivalence class of finite lists of set-theoretic
partial functions. Others were more complex than those. None of these are themselves
set-theoretic functions although they involve such functions.

Philma: “nki006” — 2005/1/21 — 10:31 — page 50 — #7

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 51

As I said in my paper [2004], there is no similar variety of uses of

axioms for set theory, but ‘the “category axioms”, as given by Eilenberg
and Mac Lane in 1945, are used in myriad ways to myriad ends in the
daily practice of mathematics’ (p. 42). Most often, so far as notices in
Mathematical Reviews show, they describe specific categories given within
a larger foundation. Rather often they are used ‘algebraically’ in general
category theory or its branches. There are other technical uses which would
normally not even be remarked in Mathematical Reviews such as when a
poset P is treated as a category, so that ordered systems indexed by P can

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

be treated as functors. And just sometimes they describe one or another
categorical foundation. These different uses have wholly different kinds of
presuppositions.
We are concerned with the presuppositions of various categorical found-
ations. I discussed several in my paper, but here I shall focus on one
that Hellman discusses, my paper ‘Axiomatizing a category of categor-
ies’ [1991].8 It is based on the ‘Category of categories as a foundation’
axioms (CCAF) from Lawvere [1966]. Hellman’s discussion suggested
three questions to answer when such axioms are offered as a categorical
foundation:
(1) What concepts are presupposed in such an axiomatization?
(2) Do these sustain the autonomy of category theory vis-à-vis set
theory, or do they reveal a (possibly hidden) dependence thereon?
(3) What is the scope of such a (meta) theory, in particular, what are
the prospects for self-applicability and the idea of ‘the category of
(absolutely) all categories’?
‘Presupposition’ in the broad Feferman-Hellman sense includes motiv-
ations. The motivation of my article was to axiomatize ‘the major theorems
of category theory’ (McLarty [1991], p. 1243). As to presuppositions in
the ordinary sense of unstated assumptions used in the axiomatization,
there are none. It is stated to be a ‘two-sorted first-order theory . . . [with]
Boolean logic’ and all the theorems are derived from the axioms in that
logic (McLarty [1999], p. 1244). There is no hidden or overt dependence
on set theory in the formal theory.
But what about the motivation? If the major theorems of category theory
are proved in set theory, and then I want to axiomatize them, is that not a
kind of dependence on set theory? Well in the first place these theorems
are not exactly proved in set theory. Their usual naïve versions are incor-
rect in set theory. They quantify over collections too large to be ZF sets,
and manipulate them too freely for Gödel-Bernays classes, and treat them
too uniformly for Grothendieck universes. There are many well-known
8 That is, Hellman discussed this in Chicago. Again, what he presented there was a work
in progress. So I do not quote it directly but acknowledge its role in raising these topics.

Philma: “nki006” — 2005/1/21 — 10:31 — page 51 — #8

52 McLARTY

and sufficiently workable set-theoretic fixes for handling these theorems

but they are all just that—fixes. The motives exceed set theory. And in
the second place even if those theorems were set-theoretically proved in
just the form I axiomatize, that would no more make my axioms ‘presup-
pose’ set theory than Hilbert’s axioms for geometry ‘presuppose’ Thales
and Pythagoras (Hilbert [1971]). In fact it would say less than that since
these theorems certainly did not come from set theory the way many of
Hilbert’s did come from Greek geometry. It would only indicate a vast
intellectual debt to earlier mathematics, including set theory, which I am

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

sure all category theorists freely acknowledge.
The key point to grasp here is precisely that categorical foundations
for category theory are not set-theoretic foundations for category theory.
When we axiomatize a metacategory of categories by the axioms CCAF,
the categories are not ‘anything satisfying the algebraic axioms of category
theory’—i.e., the Eilenberg-Mac Lane axioms. They are anything whose
existence follows from the CCAF axioms. They are precisely not sets sat-
isfying the Eilenberg-Mac Lane axioms. They are categories as described
by Lawvere’s CCAF axioms.
The third question raises two separate issues. Self-applicability is the
question of whether these axioms can be extended to posit a ‘category of
all categories’ as for example New Foundations set theory posits a set of all
sets. Not much is yet known about that and I raised the issue only briefly in
the article ([1991], p. 1243).9 But even if there is such a category it will not
be the category of absolutely all categories any more than some extension
of New Foundations posits the set of absolutely all sets.
Hellman’s first two questions apply to any new candidate foundation.
In particular they suggest a contrast which has not been emphasized up
to now between Synthetic Differential Geometry (SDG) on one hand and
Smooth Infinitesimal Analysis (SIA) on the other. Both describe categories
of spaces, including a line R and an infinitesimal subspace of it D
R, with
properties such that every function in this category has a derivative. The
axiom that guarantees this is called the Kock-Lawvere axiom. SDG consists
of the topos axioms, the Kock-Lawvere axiom, and possibly further axioms
to strengthen the theory. SIA is John Bell’s term for a theory omitting some
of the topos axioms, though also open to further stronger axioms (Bell
[1998]).
Bell uses SIA exactly to leave the logical presuppositions somewhat
open, unfixed, to suit his genetic/pedagogical account of the subject. Bell
relies on the fact that the essential analytic or geometric ideas of the subject
do not depend on the topos machinery. But the topos axioms do something
else: They allow us to conceive of SDG as a foundation because they answer
Hellman’s question 1, and so permit an answer to his question 2.

9 New Foundations indeed proves there is a ‘category of all categories’ but with
hopelessly bad properties. See McLarty [1992].

Philma: “nki006” — 2005/1/21 — 10:31 — page 52 — #9

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 53

What concepts are presupposed by the Kock-Lawvere axiom? In SDG

we presuppose standard topos logic—or, in foundational accounts, we
state that the context is the topos logic axioms and presuppose nothing.
In SIA the presuppositions are meant to be less clear. Taken this way, SIA
works rather as Awodey ([2004], pp. 58 ff.) says most mathematics does,
not specifying the background assumptions or ‘ambient theory’. That is
a fine way to work for some purposes but Hellman is right that we also
have foundational concerns. When we pursue those we cannot be satis-
fied with Awodey’s equation, where he says ‘the question of whether the

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

conditions [for a given theorem] are ever satisfied’ is just the question
of ‘whether they are consistent’ ([2004], p. 60). Different logical presup-
positions make different theories consistent.10 On Awodey’s view (p. 62)
when stronger existence assumptions are used this merely means ‘specify-
ing more of the ambient structure to be taken into account’. I am very
sympathetic to that. But it does posit an ambient structure and another
indispensable part of mathematics (and not only the philosophy of math-
ematics) is the effort to articulate the ambient structure for any body of
work.
This brings us to the other key question in Hellman’s paper: How can
a structuralist theory pick out any intended interpretation? If we identify
‘structuralist’ theories with ‘algebraic’ theories in the sense of theories
that describe only a general structure with many different instantiations,
then they cannot pick out any specific model. On my view categorical
foundations are not structuralist in that sense. Each one posits a specific
category, and I quoted Mac Lane and Lawvere expressing this view of the
category of sets (McLarty [2004], pp. 43–44).11 They are structuralist in
this precise sense: They attribute only structural properties to their objects,
that is only isomorphism-invariant properties.
This goes to the impetus behind most of the recent interest in
‘structuralist’ conceptions of mathematics. Benacerraf argued that numbers
cannot be sets because, for example, the set Zn of Zermelo natural
numbers has
1 = {∅}, 2 = {{∅}}, 3 = {{{∅}}}, . . .

while the set V n of von Neumann natural numbers has

1 = {∅}, 2 = {∅, {∅}}, 3 = {∅, {∅}, {∅, {∅}}}, . . .

10 As a trivial example, x ∈ x is consistent with some membership-based set theories and

inconsistent with others. The SIA axioms are consistent in topos logic but inconsistent with
the law of excluded middle.
11 Hellman warns against pretending to all embracing completeness in foundations. I
agree. I only find it a less pressing issue because I would not know how to pretend to it if
I tried. Clearly there are more sets, categories, smooth spaces, or whatever, than any given
axiomatization can prove, and no axiomatization I have ever seen denies it.

Philma: “nki006” — 2005/1/21 — 10:31 — page 53 — #10

54 McLARTY

The two sets V n and Zn both model the Peano axioms for arithmetic
and so both have equally good claim to be the natural numbers. They are
isomorphic and indeed any set isomorphic to them can represent the natural
numbers. Yet they have different properties, and so Benacerraf says they
cannot both be the natural numbers, thus neither one can, nor can any other
set be the natural numbers (Benacerraf [1965], p. 57). For one explicit
example, let P (X) be the ZF formula

P (X) = (∀Y ∈ X)(∅ = Y ∨ ∅ ∈ Y ).

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

In words that says every nonempty member of X has the empty set as a
member. Then ZF proves:

P (V n) and ¬P (Zn).

For technical purposes notice P (X) is equivalent to a formula with no

constants and no free variables but X:

P (X) = (∀Y ∈ X)((∀Z)(¬Z ∈ Y ) ∨ (∃W ∈ Y )(∀Z)(¬Z ∈ W )).

So Benacerraf asks for a theory of structures in which:

numbers are not objects at all, because in giving the proper-
ties (that is, necessary and sufficient) of numbers you merely
characterize an abstract structure—and the distinction lies in
the fact that the ‘elements’ of the structure have no properties
other than those relating them to other ‘elements’ of the same
structure. (Benacerraf [1965], p. 70)
The sets of categorical set theory are themselves abstract structures in
exactly this sense. An element x ∈ S in categorical set theory has no prop-
erties except that it is an element of S and is distinct from any other elements
of S. This is discussed historically and philosophically in Lawvere [1994].
Benacerraf’s goal was met in the Proceedings of the National Academy of
Science one year before Benacerraf posed it (Lawvere [1964]).
In categorical set theory isomorphic sets S ∼ = S  provably have all
the same properties. To be quite explicit: Let P (X) be any formula in
categorical set theory, with no constants and no occurrence of variables
S or S  . Let Isom(S, S  ) be the formula saying S and S  are isomorphic.
Then the following statement is provable. It is a theorem of categorical set
theory:
Isom(S, S  ) ⇒ [P (S) ⇔ P (S  )].
For details see McLarty [1993]. Categorical set theory can express no
properties that can distinguish between isomorphic sets. The usual versions

Philma: “nki006” — 2005/1/21 — 10:31 — page 54 — #11

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 55

cannot even prove there are distinct isomorphic sets. They are consistent
with a skeletal axiom saying:12

Isom(S, S  ) ⇒ S = S  .

In categorical foundations for category theory, such as the CCAF

axioms, isomorphic categories C ∼ = C  provably have all the same proper-
ties. This is normal for categorical foundations as it is for the categorical

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

methods mathematicians use every day in practice.
Altogether, I think Hellman has asked the right questions about
foundations, and categorical foundations answer them.

3. Historical Appendix on Linear Transformations Prior to

Vector Spaces
Steve Awodey and Dana Scott in Chicago mentioned two senses in which
the derivative was known as a linear transformation before any vector
spaces were defined for it to transform. Long before anyone conceived of
infinite-dimensional vector spaces, or vector spaces of functions, it was
known that taking derivatives preserves addition and real multiplication.
That is, for any functions f, g :R→R and any real number a ∈ R, using a
prime to indicate the derivative:

(f + g) = f  + g  and (a · f ) = a · f  .

This property of the derivative was used from the beginning to evaluate
derivatives and integrals, and especially to solve what were called ‘linear’
differential equations at least as early as 1853 (Petzval [1853]).13 The
derivative was soon seen as a linear transformation on functions and ‘by
the late nineteenth century it was apparent that many domains of mathem-
atics dealt with transformations or operators acting on functions’ (Kline
[1972], p. 1076). But these transformations were understood just that
12 I never use skeletal axioms because I believe they achieve nothing of interest. Take
the example of the natural numbers. Categorical set theory can assume there is exactly one
set N modelling the Peano axioms, but there are still provably infinitely many different
models. They differ in the choice of a zero element 0 ∈ N and successor function N→N.
Given any model N, 0, s :N→N, there are provably infinitely many (recursive) non-identity
isomorphisms u :N→N, and for each of them there is another model of the Peano axioms
N, u0, u−1 su :N→N. The skeletal axiom does not eliminate multiple models. Even without
a skeletal axiom, categorical set theory already eliminates any distinction among the multiple
models.
13 For one important historical example, see how George Hill studied the moon’s motion
by treating certain differential equations as infinite systems of infinite linear equations,
ridiculed by his contemporaries until Poincaré took it up (Kline [1972], pp. 731 f.).

Philma: “nki006” — 2005/1/21 — 10:31 — page 55 — #12

56 McLARTY

way: as operating on functions and not as transforming vector spaces of

functions. The theory was well advanced decades before anyone formu-
lated the idea of any space of functions, or of any infinite-dimensional
vector space.
Today there are many different ways to formalize these ideas. For
example the set C ∞ (R, R) of smooth real-valued functions on the reals
forms an infinite-dimensional real vector space, and there is a linear
transformation
C ∞ (R, R) → C ∞ (R, R)

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

taking each function f to its derivative f  . Those ideas seem to begin with
Fréchet around 1906 (Kline [1972], p. 1078).
It has also been known essentially from the beginnings of calculus that
the derivative fx of a function f at a point x is a kind of linear transformation
around x. For the case of a real-valued function f :R→R on the reals the
derivative fx was early seen as the slope of the tangent line to the graph
of f , at the point (x, f (x)). The tangent is a linear approximation to the
graph. This was soon construed as a Taylor series expansion of f around
x. Taking x as fixed, for any real number y ∈ R:

f (x + y) = f (x) + y · fx + O 2 (y),

where O 2 (y) is some function of y that vanishes to second order. Then,

either the second term y · fx or the first two terms f (x) + y · fx can be seen
as a linear function of y. The terminology was unsettled in the nineteenth
century.
Deeper ideas applied to functions f :M → N between differentiable
manifolds. It was known, and crucial, that such a function has a derivative
fx at each point x ∈ M, and fx is a linearization of f , but in what sense?
Intuitively, fx is linear from the infinitesimals around x ∈ M to those
around f (x) ∈ N. But infinitesimals were not too precisely defined, and
few if any nineteenth-century mathematicians understood infinitesimals as
forming anything like vector spaces around points. For calculation, one
could put co-ordinate systems around x and f (x) and represent fx as a
linear function of those co-ordinates pretty much the way f (x)+y·fx was a
linear function of y in the case of a real-valued function on the reals.
But fx existed independently of any co-ordinate systems. It was decades
before differential geometers articulated the idea of the tangent space to a
manifold at a point.14 Then the derivative fx :Mx →Nf (x) became a linear

14 Steve Awodey specifically mentioned the tangent bundle TM to a manifold M. This

combines all the tangent spaces to M in a coherent way so that the derivatives at all points
form a single function f  :TM→TN, which is linear in a more sophisticated sense.

Philma: “nki006” — 2005/1/21 — 10:31 — page 56 — #13

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 57

transformation from the tangent space Mx of M at x, to the tangent space

Nf (x) of N at f (x).
In both cases mathematicians used general theories of certain kinds of
specific linear transformations, which they explicitly regarded as linear
transformations, before they had any account of the linear spaces those
transformations would transform. In both cases it was real progress to
articulate spaces for the transformations: in the first case to articulate
infinite-dimensional function spaces, and in the second to articulate tan-
gent spaces and tangent bundles. The axiomatization of Abelian categories

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

similarly led to substantial progress in topology, algebraic geometry, and
number theory.

4. Logical Appendix on Linear Transformations Prior to

Vector Spaces
Here is a short account of how to pass directly from the historical starting
point of linear algebra, using vectors and matrices as calculational tools,
to the current theory of vector spaces and linear transformations, without
saying anything about vector spaces except that they are the termini of
linear transformations. This is not at all deep, does not use the Abelian
category axioms, and need not rest on categorical foundations.
Take any set-theoretic foundation whether it be naïve Cantorian set
theory, Zermelo-Fraenkel, or categorical set theory. In any case define a
category of matrices this way:
Objects are natural numbers n.
An arrow f :n→m is an m × n matrix of real numbers.
Arrows compose by the familiar matrix multiplication.
Clearly this is equivalent to the conventional category of finite-
dimensional real vector spaces and linear functions. The arrows to and from
any natural number n are exactly the conventional real-linear functions to
and from the conventional vector space Rn .
Define a vector v in an object n to be an arrow v :1→n. That means it is
an n-tuple of real numbers r1 , r2 , . . . rn , treated as the column of a n × 1
matrix. Any arrow f :n→m takes each vector v in n to a vector f (v) in m,
namely the composite
v f
f (v) = 1 → n → m

which by definition is the usual matrix operation on a column vector. This

notation makes our category look just like the conventional category of
finite-dimensional real vector spaces. The conceptual difference remains
that, instead of being elements of vector spaces, vectors are arrows to
vector spaces. The coordinate-free methods of linear algebra appear in this
category as the usual ‘up to isomorphism’ categorical methods.

Philma: “nki006” — 2005/1/21 — 10:31 — page 57 — #14

58 McLARTY

Infinite-dimensional vector spaces use infinite matrices: For any sets S

and T , define a T × S real matrix to be a set of real numbers Mx,y for
each x ∈ T , y ∈ S such that, for each y ∈ S there are only finitely many
x ∈ T with My,x = 0. These are the usual matrices when S and T are
finite. Define the product of a U × T matrix N with a T × S one M by the
obvious formula:

(N · M)x,z = Mx,y Ny,z .
y∈T

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

The sum is well defined since all but finitely many terms are 0.
Then define a category whose objects are sets S, and arrows M :S→T are
T × S real matrices composing by matrix multiplication. This is equivalent
to the usual category of all real vector spaces. For vector spaces over any
field k simply use matrices of elements of k.
We do not only want linear algebra. We want to apply calculus, say,
to smooth paths in finite-dimensional real vector spaces. We cannot define
smooth functions c :R →n for a ‘vector space’ n. We can define smooth
functions c :R→Hom(1, n), in the usual way since Hom(1, n) is the set of
n-tuples of real numbers, and then everything goes just as in undergraduate
textbooks.
Such smooth functions to linear hom sets are no kind of makeshift
or detour. Whatever definitions we use they are indispensible in dynam-
ics, Lie group theory, and elsewhere. We can use either our matrix hom
sets Hom(n, m) or the conventional vector space ones HomR (Rn , Rm ).
Either way an n-dimensional linear family of smooth trajectories in
m-dimensional space is a smooth function

t t
R → Hom(n, m), R → HomR (Rn , Rm ).

Looked at another way, t represents a smooth one-parameter family of

linear transformations from n-dimensional to m-dimensional space.
This approach to linear transformations hardly changes anything from
the current textbook approach. It is just one trivial way to define linear
transformations without defining or describing linear spaces except to say
they are domains and codomains of the transformations.

References
Awodey, Steven [2004]: ‘An answer to G. Hellman’s question: “does category
theory provide a framework for mathematical structuralism?” ’, Philosophia
Mathematica (3) 12, 54–64.
Bell, John, L. [1981]: ‘Category theory and the foundations of mathematics’,
British Journal for the Philosophy of Science 32, 349–358.

Philma: “nki006” — 2005/1/21 — 10:31 — page 58 — #15

 LEARNING FROM QUESTIONS ON CATEGORICAL FOUNDATIONS 59

—— [1986]: ‘From absolute to local mathematics’, Synthese 69, 409–426.

—— [1988]: Toposes and Local Set Theories. Oxford: Oxford University Press.
—— [1998]: A Primer Of Infinitesimal Analysis. Cambridge: Cambridge Univ-
ersity Press.
Benacerraf, Paul [1965]: ‘What numbers could not be’, Philosophical Review
74, 47–73.
Dieudonné, Jean [1981]: History of functional analysis. Amsterdam: North-
Holland.
—— [1989]: History of Algebraic and Differential Topology, 1900–1960.

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

Basel, Boston: Birkhäuser.
Eilenberg, Samuel, and Saunders Mac Lane [1945]: ‘General theory of
natural equivalences’, Transactions of the American Mathematical Society
58, 231–294.
Feferman, Solomon [1977]: ‘Categorical foundations and foundations of
category theory’, in R. Butts and J. Hintikka, eds., Logic, Foundations of
Mathematics, and Computability Theory. Dordrecht: D. Reidel, pp. 149–169.
Freyd, Peter [1964]: Abelian Categories: An Introduction to the Theory of
Functors. New York: Harper and Row.
Grothendieck, Alexander [1957]: ‘Sur quelques points d’algèbre homo-
logique’, Tôhoku Mathematical Journal 9, 119–221.
Hartshorne, Robin [1977]: Algebraic Geometry. Berlin: Springer-Verlag.
Hellman, Geoffrey [2003]: ‘Does category theory provide a frame-
work for mathematical structuralism?’, Philosophia Mathematica (3) 11,
129–57.
—— [to appear]: ‘What is Categorical Structuralism?’, in J. van Benthem,
G. Heinzmann, M. Rebuschi and H. Visser, eds., The Age of Alternative
Logics. Assessing philosophy of logic and mathematics today. Dordrecht:
Kluwer.
Hilbert, David [1971]: Foundations of Geometry. La Salle, Ill.: Open Court
Press.
James, I. M. [1999]: History of Topology. Amsterdam: North-Holland.
Kline, Morris [1972]: Mathematical Thought from Ancient to Modern Times.
Oxford: Oxford University Press.
Lang, Serge [1965]: Algebra. Reading, Mass.: Addison-Wesley.
Lang, Serge [1993]: Algebra. Reading, Mass.: Addison-Wesley.
Lawvere, F. William [1964]: ‘An elementary theory of the category of sets’,
Proceedings of the National Academy of Science of the U.S.A 52, 1506–1511.
—— [1966]: The category of categories as a foundation for mathematics. In
Samuel Eilenberg et al., editors, Proceedings of the Conference on Categorical
Algebra, La Jolla, 1965, Berlin: Springer-Verlag, pp. 1–21.
—— [1994]: ‘Cohesive toposes and Cantor’s “lauter Einsen” ’, Philosophia
Mathematica (3) 2, 5–15.
Mac Lane, Saunders [1986]: Mathematics: Form and Function. Berlin:
Springer-Verlag.
—— [1998]: Categories for the Working Mathematician. Berlin: Springer-Verlag,
2nd edition.

Philma: “nki006” — 2005/1/21 — 10:31 — page 59 — #16

60 McLARTY

Mac Lane, S., and Garrett Birkhoff [1967]: Algebra. New York:
Macmillan.
McLarty, Colin [1991]: ‘Axiomatizing a category of categories’, Journal of
Symbolic Logic 56, 1243–1260.
—— [1992]: ‘Failure of cartesian closedness in NF’, Journal of Symbolic Logic
57, 555–556.
—— [1993]: ‘Numbers can be just what they have to’, Noûs 27, 487–498.
—— [1998]: ‘Category theory, applications to the foundations of mathematics’,
in Edward Craig, editor, Routledge Encyclopedia of Philosophy. London:

Downloaded from https://academic.oup.com/philmat/article/13/1/44/1569370 by guest on 24 November 2020

Routledge.
—— [2004]: ‘Exploring categorical structuralism’, Philosophia Mathematica (3)
12, 37–53.
Petzval, Joseph [1853]: Integration der linearen Differentialgleichungen mit
Constanten und veränderlichen Coefficienten. Wien: W. Braumüller.
Russell, Bertrand [1919]: Introduction to Mathematical Philosophy. New
York: George Allen and Unwin.
Shapiro, Stewart [1991]: Foundations without foundationalism: A case for
second-order logic. New York: Oxford University Press.
Stone, Marshall Harvey [1932]: Linear Transformations in Hilbert Space
and Their Application in Analysis. American Mathematical Society Col-
loquium Publications, Volume XV. Providence, R.I.: American Mathematical
Society.

Philma: “nki006” — 2005/1/21 — 10:31 — page 60 — #17